| Good morning. It's Wednesday, November 11th. | | | In partnership with Adobe | | Adobe Firefly, the all-in-one AI creative studio, is expanding the Adobe Creative Agent through new capabilities in Firefly AI Assistant (beta). It carries out multi-step tasks using Adobe Creative Cloud tools, including Premiere, Illustrator, InDesign, After Effects and Frame.io, all guided by your direction. Describe what you want to make in your own words, then refine, redirect, or take over at any time. | Adobe is also previewing Adobe Creative Cloud MCP, a new way for creative professionals to direct third-party agents as they make changes directly inside native project files in Photoshop, Illustrator, Premiere and After Effects, while staying in control. | It's part of Adobe's approach to making pro-grade creative workflows easier to start, control and continue, wherever inspiration strikes. | Explore what’s new from Adobe at MAX. | Thank you for supporting our sponsors! | | Anthropic priced its IPO at up to $1.4 trillion, aims for next week | Anthropic launched its IPO roadshow yesterday with real terms attached. 825 million Class A shares at $70 to $78, listing on the Nasdaq Global Select Market under ANT. Anthropic is selling 801,351,354 of those shares and existing holders are selling 23,648,646, with a 30-day underwriter option on up to 123.75 million more. At the top of the range that is up to $64.4 billion raised and roughly $1.4 trillion fully diluted, per Fortune. Goldman Sachs and Morgan Stanley are listed as the sources for the preliminary prospectus. | Seven strategic partners, reportedly including Nvidia, Microsoft, Google, TSMC, Broadcom, AMD and Micron, agreed to buy $26 billion of non-voting Class N stock at the IPO price. Because it is a private placement, that block does not carry the standard six-month lockup. The Wall Street Journal, via Yahoo Finance, puts the debut on Thursday November 19 or Friday November 20. The Financial Times reports the book is already oversubscribed, with some bankers weighing a valuation above $1.5 trillion. Treat that as reporting, not a filed number. | The Nasdaq fell 5.5% last week on post-midterm worries about AI regulation and data center scrutiny, and Anthropic's own prospectus warns that using AI to build more capable AI "could also pose a considerable safety risk," adding that internal estimates suggest it "may imminently meet" the acceleration threshold in its Responsible Scaling Policy. The amended filing also discloses that its $34.9 billion Lambda compute deal is now terminable after a missed lease deadline. We covered the original S-1 here. | Read more. | | OpenAI lost $10.9 billion last quarter on $10 billion of revenue | The Wall Street Journal, working from documents it obtained, reports OpenAI booked about $10 billion of revenue in the third quarter, up roughly 50%, against an adjusted operating loss of $10.9 billion. That is an adjusted operating margin near -109%, improved from about -150% in Q2. OpenAI is private, files nothing with the SEC, and has not commented on the figures. | The comparison everyone will reach for is Anthropic’s: Its amended S-1 from November 2 shows third-quarter revenue of about $17 billion and an operating loss of $567 million, driven by $5 billion of short-term compute commitments and $2 billion of non-cash charges tied to a donation matching program. | OpenAI's number is an adjusted figure from leaked documents; Anthropic's is GAAP in a filing with legal consequences, and the two companies draw their revenue lines differently. What it does establish is that the two largest labs are now running on very different unit economics, and only one of them has to show the math to public investors next week. | Read more. | | Claude Fable 5.2 and Mythos 5.2 have arrived | Anthropic shipped Fable 5.2 and Mythos 5.2 on Monday. Same underlying model, two different safeguard configurations: Fable carries the cyber and biology restrictions and is open to everyone, while Mythos has fewer blocks and is gated behind the Cyber and Life Sciences Verification Programs. Pricing is $8 per million input tokens and $40 output, doubling to $16 and $80 for an optional fast mode that generates up to 2.5x quicker. | The benchmark deltas against Opus 5.5 are mostly in agentic and scientific work: 69.9% on FrontierSWE v2 versus 62.3%, 74.6% on Terminal-Bench-Science versus 58.7%, 41.3% on the Remote Labor Index versus 21.3%, and 64.9% on RoboDojo Sim versus 39.4%. Worth noting what did not move: Opus 5.5 still edges Fable 5.2 on Anthropic's Elo-style knowledge work score (1800 to 1785), and Google's Gemini 4 Carbon beats Fable on AutomationBench, 53.9% to 51.4%. Anthropic says it loosened Fable's safeguards too, with roughly 55% fewer blocks than Fable 5.1. | Anthropic says a 1,000-agent team formalized Fermat's Last Theorem in about 30 hours, down from 11 days for an earlier internal run; that Fable 5.2 resolved two open problems in black hole and quantum gravity theory; that Mythos 5.2 modeled 28 of 34 attempted AAV capsids; and that over 90% of changes to Anthropic's core codebase are now written by Claude. External evaluators included the UK AI Security Institute and METR, whose acceleration assessment we covered Monday. | Read more. | | Claroc says 35,578 Grok Build apps opened without sign-in, with 650 to 800 holding private data (TeslaNorth) DeepMind's AlphaProof Code formally verified Signal's cryptographic core in about four weeks (Google DeepMind) Nvidia shareholders filed a securities class action over the Huang disclosures (CBS News) Australia will create an AI safety regulator with mandatory incident reporting after an OpenAI agent breach (Nikkei Asia) Dario Amodei will do a live-streamed IPO conversation with JPMorgan CEO Jamie Dimon (Bloomberg) Nvidia fell about 5% Monday and is down roughly 26% over the past month (TheStreet) Notion says AI chat transcripts were exposed in the Fireworks breach, affecting under 1% of workspaces (Notion) A Railway breach ran up $740,000 in fraudulent AI charges after attackers took provider credentials (BleepingComputer) CISA added a CVSS 9.8 authentication flaw in GPUStack to its known-exploited catalog (Cloud Security Alliance) Stanford's 2026 AI100 report argues scaling is hitting data and compute limits and shifting toward reasoning (InfoDocket) OpenAI acquired Odessia, the AI travel agent startup from Sonder's founder (TechCrunch) SpaceX plans a $24 billion Starbase supercompute project paired with a 748MW gas plant (Crypto Briefing) Hassabis says society is not ready, and calls closed-loop self-improvement with no human in the loop very dangerous (Masters of Scale) 38% use generative AI for information weekly, but only 9% use it for news (Nieman Lab) FraudBench finds LLM fraud rates climb from 3.0% as an advisor to 21.4% as an agent (Hugging Face) Meituan open-weighted LongCat-2.5, a 1.6-trillion-parameter MoE with vision, under MIT (VentureBeat) Adobe will open Photoshop and Premiere to third-party AI agents over MCP in an early 2027 private beta (PYMNTS) Musk says Grok 4.8 soft-rolls out this week after a rebuild that swapped JAX for C++ (TeslaNorth) Daylight raised an $80 million Series B led by Goldman Sachs Alternatives for AI security operations (Pulse 2.0) PM job openings hit a four-year high while software engineering hiring sits below half its 2022 peak (Lenny's Newsletter) Google launched a US preview of Gemini for Healthcare with seven partners including Northwell, HCA and Optum (Fortune) Catching up from Monday: the Nvidia board investigation, and OpenAI's $11 billion first tranche (AI Breakfast)
| | Gemini 4 Carbon went generally available yesterday, two weeks after it was announced. Google calls it its strongest model and is explicit that most people should not use it: Gemini 4 Argon remains the recommendation for ordinary software engineering and knowledge work, because Carbon burns far more tokens reasoning and verifying. Introductory pricing matches Argon at $2 input and $10 output per million tokens, then rises to $4 and $20 on November 30. | In robotics, Nvidia used CoRL to open up GR00T 2, a world action model on Cosmos that hit 71% average success across 50 manipulation tasks with one demonstration each, rising to 95% with five, plus Isaac Lab 3.0 with switchable PhysX and Newton solvers and up to 2.5x the training throughput of 2.3. Skild AI open-sourced Steel, a Newton-based physics solver it built with Nvidia, claiming 3.8x PhysX throughput across 20 benchmarks. And IBM and Microsoft announced Imagine Enterprise with roughly 6,000 pre-built industry agents, a number that tells you more about packaging than capability. | | Thank you for reading today's edition. | | Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter. | Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X! | Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months. |
|